Long-horizon robot manipulation is a challenging control problem because the policy must not only execute individual actions but also coordinate a sequence of intermediate subgoals and decide when to advance from one to the next. The authors of this paper argue that most imitation learning methods are trained solely on task-level objectives, meaning they never explicitly learn the subgoal structure that governs long-horizon tasks.

To address this, they introduce SADP (Subgoal-Aware Diffusion Policy), a diffusion-based policy that is explicitly aware of subgoals. The policy is trained on demonstrations generated by foundation models, which likely provide the subgoal annotations needed for this learning signal. By incorporating subgoal awareness, the policy is designed to better handle the timing of transitions between subgoals, a key requirement for long-horizon manipulation.

The paper is available as a revised arXiv preprint (2605.16871v3). The provided abstract focuses on the problem formulation and the proposed method; it does not include quantitative results in the excerpt. Still, the work highlights a clear gap in current imitation learning approaches and offers a concrete way to close it.